Date: Sun, 1 Apr 2012 13:38:57 +0400 From: Solar Designer <solar@...nwall.com> To: john-dev@...ts.openwall.com Subject: CUDA vs. OpenCL; approach to fast hashes (was: rawsha256.cu patch(using shared memory)) On Thu, Mar 29, 2012 at 05:16:21AM +0800, myrice wrote: > I am sorry for lack my hardware details. GTX-580 is my lab's server. But > recently it becomes unstable :( > I tested this code on my laptop with GeForce 9600M GS card and P8600 CPU. > So the performance is slow. Do you expect to have access to faster cards during the summer? Do you intend to write any OpenCL code or only CUDA? Efficient OpenCL code is more desirable than CUDA code because AMD cards are simply about 3 times faster than Nvidia at these tasks. This is going to be partially corrected with GTX 680, but I expect a difference of about 1.5x in favor of AMD to remain (for optimized code). And that's per-chip. Now if we consider that there are fast dual-GPU AMD cards on the market (6990, 5970), but no dual-GPU counterpart to GTX 680 available yet, that gives AMD even further advantage this year (more GPUs per machine). Perhaps a new dual-GPU card from Nvidia will appear soon, though. > However, I think If I implement candidate password generation and > comparison on GPU, there are lots of work to do. I have to go > through existing code on password generation(I guess they are mainly in > Crakc.c?) and subtitute it with cuda. That's one way to do it, but not the only one. While there's obvious usability advantage from being able to have all the same candidate passwords generated on GPU that would be generated on CPU, it also makes sense to have special cracking modes that are more suited to offloading to GPU. And no, this is not in cracker.c, but in per-cracking-mode source files. Alexander
Powered by blists - more mailing lists
Confused about mailing lists and their use? Read about mailing lists on Wikipedia and check out these guidelines on proper formatting of your messages.